De-embedding
Removing the test fixture's own response from a measurement so that only the device under test remains. In photonics this usually means subtracting the fiber-to-chip couplers, measured on a dedicated reference structure, from every on-chip device measurement.
No fiber-coupled measurement of a photonic chip measures the device alone. Light enters through a coupler, crosses the device, and leaves through another coupler, so the raw number is always
and the couplers are rarely small: a grating coupler costs several dB and, worse, has a strongly wavelength-dependent, roughly Gaussian passband that imprints itself on every spectrum measured through it. De-embedding is the removal of those fixture terms.
The reference structure. The standard fixture calibration is a back-to-back (loopback) structure on the same die: two nominally identical couplers connected by a short waveguide and nothing else. Its measured loss, minus the negligible waveguide term, is two couplers' worth, so the per-coupler insertion loss spectrum is half the loopback loss in dB. Subtracting two of these curves from a device path leaves the device. Because subtraction in dB is division in linear power, spectral de-embedding is a division of transmission spectra, which is why the reference must be measured over the same wavelength grid, same polarization, and ideally the same fiber landing.
The assumption doing all the work is that the reference couplers equal the device-path couplers. Across one die that is usually good to a few tenths of a dB; across a wafer, thickness and linewidth gradients shift both the peak loss and the center wavelength of the coupler passband, which is why wafer-level test flows place a loopback in every reticle rather than trusting one per wafer. Alignment repeatability sets the floor: if repeated landings on the same coupler scatter by 0.1 dB, no amount of arithmetic gets device numbers better than that. Measure the scatter once by landing the probe several times, and quote it.
Ripple is a diagnostic, not just noise. Periodic wiggles on a de-embedded (or raw) spectrum usually mean a parasitic Fabry-Perot cavity: reflections at the two couplers form a resonator whose fringe spacing obeys . Read the period, and with the group index you get the cavity length , which usually points straight at the culprit: coupler-to-coupler distance means coupler back-reflection, facet-to-facet distance means the chip edges. The ripple's peak-to-peak amplitude bounds the reflectivities.
Beyond couplers. The same logic runs through all of photonic metrology: the cutback method de-embeds coupling by comparing structures that differ only in length; RF characterization de-embeds probes and pads with short/open/thru structures; detector calibrations de-embed the photodiode's spectral response. The shared discipline is that every measurement includes its fixture, and only a deliberately designed reference lets you take the fixture back out.
References: Chrostowski & Hochberg, Silicon Photonics Design (2015), Ch. 9; the companion procedure with worked numbers is De-embedding coupler loss with reference structures, and the Grating Coupler Spectrum Fitter extracts per-coupler loss, bandwidth, and ripple from a pasted loopback scan.